Google Disabled an AI Feature in Earth After a Dipfake Incident

Google disabled the image-generation feature in Google Earth shortly after its launch: users had been able to create realistic scenes of disasters, military installations, and destruction based on satellite data. The problem here isn’t just a lack of effective safeguards—the very status of satellite imagery as relatively reliable visual evidence has come under fire.

The feature used Nano Banana 2 and was available to regular users without special access or a waiting list. Google cited education, real estate, environmental research, and urban planning as examples of useful applications.

But that same interface allowed users to generate images of refugees at the border, damaged structures, fires, and other events that might not actually exist. Researchers and journalists quickly demonstrated that the safety restrictions could be circumvented using standard prompt engineering techniques. In the tests described, they even managed to generate images of the Eiffel Tower collapsing and a sinkhole beneath the Great Pyramid of Giza.

The main flaw was apparent even before launch

Combining a generative model with real geographic data means giving it not just the ability to draw, but the ability to fabricate context. The user gets realistic shadows, terrain, roofs, and surroundings, while social media viewers see an image that looks like a satellite photo.

This is a predictable conflict between a “creative tool” and a tool for disinformation. It didn’t need to be discovered after the first posts: it would have been enough to test scenarios involving military facilities, natural disasters, and humanitarian crises in advance.

Ultimately, Google Earth may not be the only one to suffer. Satellite imagery is used by journalists, OSINT researchers, and human rights organizations to verify destruction, troop movements, and environmental damage. If a fake image becomes visually indistinguishable from a real one, the next step isn’t necessarily exposing the forgery, but rather distrusting all images across the board.

This creates a fertile ground for propaganda: genuine evidence can be dismissed as AI-generated fakes, while fabricated images can be circulated as confirmation of a desired narrative.

Invisible protection doesn’t work as a deterrent

Google was counting, among other things, on SynthID—an invisible watermark embedded in the pixels of a generated image. But a watermark that humans can’t see doesn’t stop the image from being published. All it takes is a screenshot, saving the file, or uploading it to another platform where no one will run an origin check.

Even if the metadata and watermark are preserved, the whole system assumes that the recipient will first check the image and only then share it. On social media, the opposite usually happens: an image first evokes an emotion, then it’s shared, and only occasionally does someone start to investigate its source.

Provenance is useful for investigations, but it is no substitute for product-level security. If a dangerous result can be obtained with just two clicks, a content provenance system should not be the sole line of defense.

What to Do with Such Images

For the average reader, the rule has become uncomfortably simple: a single striking satellite image is no longer evidence in and of itself.

• Do not forward images of disasters, attacks, or crises immediately after viewing them.
• Verify the source, author, and caption of the image.
• Look for independent reports from media outlets, official agencies, and eyewitnesses.
• Pay attention to the date, timestamps, and satellite data provider.
• Treat AI detectors as an additional indicator, not a final verdict.
• If an image cannot be reliably verified, clearly label it as unverified.

For journalists and researchers, the bar is even higher: cross-checking across multiple sources, analyzing the time the image was taken, and technical expertise are required. The habit of treating a single image as definitive proof no longer holds up to reality.

Temporarily disabling the feature fixes this specific release but does not eliminate the underlying class of problems. Similar tools will emerge anyway—from Google or its competitors. Therefore, the key question for AI products is no longer “does it have a watermark?” but “could it have been anticipated that this feature would create a new type of convincing lie, and why was it released anyway?”

Source: Hacker News - Newest: ""AI" "LLM""